Predicting Atomicity Violations in Concurrent Programs via Planning

Testing concurrent programs is more difficult than testing sequential programs due to the interleaving explosion problem: even for a fixed program input, there are numerous different runs that need to be tested to account for scheduler behaviour. Testing all such interleavings is not practical. Consequently, most effective testing algorithms attempt to generate runs that are likely to manifest bugs. Atomicity violating runs have been proposed as good candidates since a large fraction of the existing concurrency bugs result from such violations.In this paper we present a general approach to predicting atomicity violations that is based on techniques from Artificial Intelligence (AI) automated planning. We encode the dynamics of our program abstractly in terms of the properties of observed events from a successful program run. We characterize the generation of a run as a sequential planning problem with the temporally extended goal of achieving a particular pattern of atomicity violation. We have integrated our approach into the PENELOPE concurrency testing tool (Sorrentino, Farzan, & Madhusudan 2010). Initial experiments comparing the run prediction time for an implementation of our approach showed it to be comparable to PENELOPE’s static analysis approach. However, there are indications that the planning approach may scale better on longer runs. Further, unlike PENELOPE’S current approach, runs predicted by our approach all correspond to concrete runs of the system. Finally, our planning-based approach has the merit that it can easily accommodate complex atomicity violation patterns by simply modifying the planning goal. For all these reasons, the planning-based approach presented here appears to be a fruitful area for further investigation.